Beihang University

Samsung Electronics Co., Ltd.

Beijing Samsung Telecom R&d Center

Dongqing Zou

Xiaowu Chen

Qinping Zhao

Qiang Wang

Keun Woo Park

Jia Qi Li

Hyunsurk Eric Ryu

Jianwei Li

Feng Shi

Feng Ding

Weiheng Liu

Xiaogang Wang

Jingtao Xu

Deheng Qian

Ping Guo

Baek Hwan Cho

Bo Gao

Qing Ql Li

Hyunku Lee

Changwoo Shin

Kan Guo

Zongji Wang

JinMan Park

Guangying Cao

An object detection method and a neural network system for object detection are disclosed. The object detection method acquires a current frame of a sequence of frames representing an image sequence, and extracts a feature map of the current frame. The extracted feature map is pooled with information of a pooled feature map of a previous frame to thereby obtain a pooled feature map of the current frame. An object is detected from the pooled feature map of the current frame. A dynamic vision sensor (DVS) may be utilized to provide the sequence of frames. Improved object detection accuracy may be realized, particularly when object movement speed is slow.

Object detection method and system

Disclosed is a co-segmentation method and apparatus for a three-dimensional model set, which includes: obtaining a super patch set for the three-dimensional model set which includes at least two three-dimensional models, each of the three-dimensional models including at least two super patches; obtaining a consistent affinity propagation model according to a first predefined condition and a conventional affinity propagation model, the consistent affinity propagation model being constraint by the first predefined condition which is position information for at least two super patches that are in the super patch set and belong to a common three-dimensional model set; converting the consistent affinity propagation model into a consistent convergence affinity propagation model; clustering the super patch set through the consistent convergence affinity propagation model to generate a co-segmentation outcome for the three-dimensional model set. The disclosed three-dimensional model set co-segmentation method and apparatus improves consistency between three-dimensional model set co-segmentation outcomes.

Co-segmentation method and apparatus for three-dimensional model set

The disclosure provides an object detection method and apparatus based on a Dynamic Vision Sensor (DVS). The method includes the following operations of: acquiring a plurality of image frames by a DVS; and, detecting the image frames by a recurrent coherent network to acquire a candidate box for objects to be detected, wherein the recurrent coherent network comprising a frame detection network model and a candidate graph model. By using a new recurrent coherent detection network, a bounding box for an object to be detected is fast detected from the data acquired by a DVS. The detection speed is improved greatly while ensuring the detection accuracy.

Object detection method and apparatus based on dynamic vision sensor

A method and apparatus for processing a binocular disparity image are provided. A method of determining a disparity of a binocular disparity image that includes a left eye image and a right eye image includes acquiring features of a plurality of pixels of the binocular disparity image based on an event distribution of the binocular disparity image, calculating a cost matrix of matching respective pixels between the left eye image and the right eye image based on the features, and determining a disparity of each matched pair of pixels based on the cost matrix.

Method and apparatus for processing binocular disparity image

Exemplary embodiments provide an image vision processing method, device and equipment and relate to: determining parallax and depth information of event pixel points in a dual-camera frame image acquired by Dynamic Vision Sensors; determining multiple neighboring event pixel points of each non-event pixel point in the dual-camera frame image; determining, according to location information of each neighboring event pixel point of each non-event pixel point, depth information of the non-event pixel point; and performing processing according to the depth information of each pixel point in the dual-camera frame image. Since non-event pixel points are not required to participate in the matching of pixel points, even if it is difficult to distinguish between the non-event pixel points or the non-event pixel points are occluded, depth information of the non-event pixel points can be accurately determined according to the location information of neighboring event pixel points.

Image vision processing method, device and equipment

A method and apparatus for acquiring an image disparity are provided. The method may include acquiring, from dynamic vision sensors, a first image having a first view of an object and a second image having a second view of the object; calculating a cost within a preset disparity range of an event of first image and a corresponding event of the second image; calculating an intermediate disparity of the event of the first image and an intermediate disparity of the event of the second image based on the cost; determining whether the event of the first image is a matched event based on the intermediate disparity of the event of the first image and the intermediate disparity of the event of the second image; and predicting optimal disparities of all events of the first image based on an intermediate disparity of the matched event of the first imaged.

Method and apparatus for acquiring image disparity

The present invention provides a method for video matting via sparse and low-rank representation, which firstly selects frames which represent video characteristics in input video as keyframes, then trains a dictionary according to known pixels in the keyframes, next obtains a reconstruction coefficient satisfying the restriction of low-rank, sparse and non-negative according to the dictionary, and sets the non-local relationship matrix between each pixel in the input video according to the reconstruction coefficient, meanwhile sets the Laplace matrix between multiple frames, obtains a video alpha matte of the input video, according to &#x3b1; values of the known pixels of the input video and &#x3b1; values of sample points in the dictionary, the non-local relationship matrix and the Laplace matrix; and finally extracts a foreground object in the input video according to the video alpha matte, therefore improving quality of the extracted foreground object.

Method for video matting via sparse and low-rank representation

The invention provides a data driven method for transferring indoor scene layout and color style, including: preprocessing images in an indoor image data set, which includes manually labeling semantic information and layout information; obtaining indoor layout and color rules on the data set by learning algorithms; performing object-level semantic segmentation on input indoor reference image, or performing object-level and component-level segmentations using color segmentation methods, to extract layout constraints and color constraints of reference images, associating the reference images with indoor 3D scene via the semantic information; constructing a graph model for indoor reference image scene and indoor 3D scene to express indoor scene layout and color; performing similarity measurement on the indoor scene and searching for similar images in the data set to obtain an image sequence with gradient layouts from reference images to input 3D scene; performing image-sequence-guided layout and color transfer generation for indoor 3D scene.

Data driven method for transferring indoor scene layout and color style

The invention discloses a deep learning based method for three dimensional (3D) model triangular facet feature learning and classifying and an apparatus. The method includes: constructing a deep convolutional neural network (CNN) feature learning model; training the deep CNN feature learning model; extracting a feature from, and constructing a feature vector for, a 3D model triangular facet having no class label, and reconstructing a feature in the constructed feature vector using a bag-of-words algorithm; determining an output feature corresponding to the 3D model triangular facet having no class label according to the trained deep CNN feature learning model and an initial feature corresponding to the 3D model triangular facet having no class label; and performing classification. The method enhances the capability to describe 3D model triangular facets, thereby ensuring the accuracy of 3D model triangular facet feature learning and classifying results.

Deep learning based method and apparatus for three dimensional model triangular facet feature learning and classifying

A method of integrating binocular stereo video scenes with maintaining time consistency includes: propagating and extracting a contour of moving object of stereo video A; integrating and deformating of parallax between moving object and dynamic scene with time consistency; color blending of moving object and dynamic scene with time consistency where a method of median coordinate fusion is utilized. The method is simple and effective to utilize a small quantity of user interactions to successfully extract moving objects from stereo video which are same in time and as consistent as possible between left view and right view to develop multiple constraint conditions to guide the integration and deformation of parallax of moving object and dynamic scene and allow moving object to conform to the rules of perspective of dynamic scene. Moreover, the deformation result of moving object is smooth and consistent and can avoid effectively the occurrence of 'dithering' phenomenon.

Method of integrating binocular stereo video scenes with maintaining time consistency

This invention provides a method of sparse representation of contents of high-resolution video images which supports content editing and propagation. It mainly comprises five steps which are: (1) to input the original images or videos of high resolution to summarize the characteristic information of their pixels; (2) to acquire the highly sparse samples of the original images or videos through the sparse representation technique; (3) to reconstruct each pixel of the input images or videos with a small number of the original sparse samples to calculate a coefficient of reconstruction; (4) to edit and propagate the original sparse samples to yield a result of new sparse samples; (5) to generate a result of final images or videos of high resolution according to the result of sparse samples and the coefficient of reconstruction. This invention only edits and propagates the highly sparse samples rather than all the information of pixels. Thus, the memory consumption can be greatly reduced so as to possibly process the images or videos of high resolution in a very small memory space. It is very potential to be widely applied in the fields of image processing, computer vision and augmented reality technique.

Method of sparse representation of contents of high-resolution video images supporting content editing and propagation

The present invention discloses an image foreground matting method based on neighborhood and non-neighborhood smoothness priors. The method primarily comprises the steps of: interactively marking foreground points and background points; initializing &#x3b1; values of each unmarked pixel of the input image by a color sampling method, calculating confidence degree of the pixel, and admitting &#x3b1; values of pixels of which confidence degree is larger than a given threshold as known pixels; calculating data term weights, neighborhood smoothness constraint term weights and non-neighborhood smoothness constraint term weights of each pixel in the input image to construct graph patterns of all pixels of the input image; and according to &#x3b1; values of the known pixels, under the constraint of the graph patterns, solving probabilities that each pixel belongs to the foreground by minimizing the energy equation so as to obtain alpha mattes.

Image foreground matting method based on neighborhood and non-neighborhood smoothness priors

The invention discloses a method for editing propagation of video and image content based on local feature structure preservation, comprising: mapping all pixels in the input original image and/or video key frames to a selected feature space; finding K nearest neighbor pixels for each pixel according to feature vectors' Euclidean distance in the selected feature space; using Locally Linear Embedding (LLE) dimension reduction to construct the locally linear relationship between each pixel and its K nearest neighbor pixels in the selected feature space; According to the present invention, it is possible to accurately perform such image or video processing as automatic color transformation, interactive color editing, gray image colorization, video cloning and image matting.

Growing community of inventors

Beijing, China

Dongqing Zou